How Our Team Won the 2023 Huawei Software Elite Challenge: Strategies, Scheduling, and Pathfinding

On April 23, 2023 the 9th Huawei Software Elite Challenge “Planck Project” global finals concluded, with 645 universities, 3,887 teams and 23,078 students competing; the team from University of Electronic Science and Technology of China won the championship.

The competition tasks simulated multi‑robot environments from Huawei Cloud AI robots, requiring participants to control robots to maximize profit through scheduling, path planning, and dynamic obstacle avoidance, covering optimal scheduling, multi‑robot coordination, and dynamic avoidance algorithms.

In the preliminary round the team initially used a generic scheduling algorithm, but realized that fitting each public dataset yielded higher scores; they refined a dataset‑specific algorithm which later became crucial in the semi‑final and final.

During the semi‑final, after adding obstacles, they fixed a critical bug in the path‑finding module that mis‑identified reachable workstations, enabling them to secure first place nationwide.

For the final, they refactored the semi‑final code to improve modularity and added flexible strategies for four map types, focusing on a universal “follow‑enemy” attack while keeping production robots active.

Overall Solution

The solution is presented top‑down: high‑level combat strategy, mid‑level scheduling, path‑finding, yielding, and low‑level motion control.

Scheduling Algorithm

Each idle robot receives a buy‑sell task; tasks are locked at the product slot of the buying workstation and the material slot of the selling workstation. The algorithm selects the task with the highest profit‑to‑time ratio, estimating profit from direct sale, unblocking benefits, and downstream production gains.

The profit components include direct_profit, unblocking_profit, and sell_profit, the latter computed by the estimate_product_profit() function.

Path‑Finding Algorithm

The map is discretized into 0.25 m × 0.25 m cells; navigation points are grid intersections. Each workstation is represented by four points within 0.35 m of its center. Robots compute the shortest path in eight directions each frame and move toward the farthest reachable point along that path.

Yielding Algorithm

Collision detection predicts imminent collisions; lower‑priority robots perform a BFS to find a safe avoidance point. If no point is found, the robot’s priority is increased to let the other robot yield.

Motion Control

Using basic physics, the desired speed is set to sqrt(2 a s) to ensure the robot can brake to a stop at the target distance s with maximum acceleration a. Angular speed is set similarly with angular acceleration α. If the heading error exceeds a threshold, the robot rotates in place before moving.

Conclusion

Despite a rough start, a serious bug before the semi‑final, and low rankings in practice matches, the team’s emphasis on a generic, adaptable algorithm allowed them to dominate the final rounds and claim the overall championship.